Data-Driven Dose-Volume Histogram Prediction.

Area of Science:

• Radiation oncology
• Medical physics
• Data science in healthcare

Background:

• Predicting dose-volume histograms (DVHs) from prior radiation therapy data is crucial for treatment planning and quality assurance.
• Existing model- and data-driven methods face limitations, particularly with smaller datasets.

Purpose of the Study:

• To evaluate the accuracy of predicting clinical DVHs using a database of prior radiation therapy cases.
• To assess the influence of organ-at-risk (OAR) to planning-target-volume (PTV) geometry on prediction accuracy.

Main Methods:

• Constructed a radiation therapy database (Oncospace) with patient data for advanced lung cancer.
• Queried DVH data for lungs, esophagus, and heart based on OAR-to-PTV geometric similarity using overlap volume histograms (OVH).
• Evaluated prediction accuracy by comparing database-derived DVHs (DVH-DB) with clinically delivered DVHs (DVH-CL) across varying numbers of similar patients.

Main Results:

• Excellent DVH prediction accuracy was achieved for lungs (interquartile range <4%) and body (IQR <1%) within 2 cm of the PTV.
• Lung DVH prediction showed high accuracy, with differences between DVH-DB and DVH-CL <3% in 14/23 patients.
• Weaker DVH prediction was observed for the esophagus and heart, with mean differences >10% in over half of the cases, indicating planning preference influence.

Conclusions:

• Prior radiation therapy data can accurately predict clinical DVHs for organs with significant tumor overlap, like the lungs.
• Prediction accuracy is limited for organs at risk (OARs) without direct tumor overlap, where treatment planning choices dominate.
• The study underscores the importance of geometric relationships and data set size in the reliability of DVH prediction models.